Scalable machine learning solution for commercial scale three dimensional geophysical inversions

Scalable machine learning solution for commercial scale three dimensional geophysical inversions

Souvik Mukherjee¹, Santi Adavani^2*, Alan Morgan³, William N. Barkhouse⁴, Ronald S. Bell⁴, Peter G. Lelievre⁵, Colin G. Farquharson⁶

¹EmPact-AI, USA, ^2*RocketML, now at S2 Labs, USA, ³Bell Geospace, USA, ⁴Drone Geoscience, USA, ⁵Mt. Allison University, Canada, ⁶Memorial University, Canada

Abstract

Application of artificial intelligence (AI) and machine learning (ML) based workflows and methodologies for geophysical data processing, imaging, and interpretation are active focus areas in industry and academia. While much progress has been made to demonstrate applicability in many use cases, key bottle neck for widespread commercial use has been the prohibitively high computational cost involved in applying the method for large scale three dimensional inverse problems.

Key changes to the form of the simulated input data used for training and the corresponding design of the architecture of the hidden layers enable approximately O(n) (where n is the number of layers in the network) reduction in the computational complexity of the training architecture. Combined with multi-GPU Distributed Deep Learning (DDL) algorithms optimized specifically for training large scale ML data, this results in significant improvements in resolution of inversion results relative to conventional least squares imaging, while computational efficiency improves by order of magnitude compared to several commonly used open-source ML architectures and platforms.

When deployed for inversion of dense, closely spaced high resolution handheld magnetometer data collected over a buried pipe in a field in Texas, the resolved three-dimensional geometry and location using the new algorithm showed over 6-fold improvement compared to conventional three-dimensional least squares inversion. When applied to an 18-fold larger data set collected by a drone-based magnetometer over a field in California, the buried complex metallic pipe like structure was resolved using little over 2 days of compute time. Similar exercise undertaken in google collab GPU platform using state-of-the-art google tensorflow would have taken 3 – 6 months to complete, suggesting a 50 – 100-fold improvement in computational efficiency.

The method was also benchmarked against Los Alamos National Laboratory’s (LANL) open-source seismic full waveform inversion (FWI) dataset. LANL trained 24000 seismic data sets simulated from various 2D velocity models using 32 P100 Tesla GPU machines in 2 hours. When inferenced on 6000 previously unseen test models, the root mean square error (RMSE) in the inverted normalized velocity models was 0.018. The current workflow on the same data set achieved a comparable RMSE of 0.012 on 6000 unseen test models after training 24000 models in 50 minutes using just 4 GPU (V100) machines, achieving nearly 20-fold improvement in computational efficiency.

In addition to magnetic and seismic data, the method is being developed for applications to electromagnetic and full tensor gravity gradiometer (FTG) data. Given the significant improvements in resolution and computational efficiency, it is expected that successful ground truth based field trials of AI based geophysical data inversion has the potential to unlock several new application areas while dramatically improving the business impact of such applications in existing ones.